DeepScence AI Tool Maps Senescent Cells with Unprecedented Accuracy
A new deep-learning method identifies aging 'zombie' cells in single-cell and spatial transcriptomics data, outperforming all existing approaches.
Summary
Researchers at Duke University developed DeepScence, an unsupervised deep-learning tool built on an autoencoder architecture, to detect senescent cells in single-cell RNA sequencing and spatial transcriptomics data. The team first created CoreScence, a curated gene set of 39 genes consistently reported across at least five published senescence gene databases, addressing the massive disagreement among existing gene sets. DeepScence uses CoreScence as input and learns complex, nonlinear gene expression patterns to score cells on a senescence continuum. Tested across in vitro and in vivo datasets from multiple platforms, DeepScence substantially outperformed existing scoring methods and the supervised SVM-based SenCID tool, achieving AUROCs exceeding 0.9 across all datasets and generalizing across species, tissues, and senescence induction contexts.
Detailed Summary
Cellular senescence — the state in which cells permanently stop dividing but remain metabolically active — is a central driver of aging and age-related diseases including osteoarthritis, pulmonary fibrosis, and Alzheimer's disease. Because senescent cells (SnCs) are rare in tissues and molecularly heterogeneous, reliably identifying them in genomic data has remained a major challenge. Prior methods either relied on single marker genes (vulnerable to dropout noise), gene-set scoring approaches that ignore nonlinear relationships, or supervised machine learning trained exclusively on in vitro bulk RNA-seq data with limited in vivo generalizability.
The researchers began by surveying nine published senescence gene sets (SnGs) and documenting alarming inconsistency: Jaccard overlap indices between pairs were below 0.2 in nearly all cases, 69% of the 2,966 genes identified appeared in only one gene set, and just 39 genes (1.3%) were reported by five or more sets. To resolve this, they constructed CoreScence — a consensus gene set of those 39 consistently reported genes, including canonical markers CDKN1A (p21) and CDKN2A (p16). Validation in bulk RNA-seq datasets confirmed that genes present in more gene sets showed stronger differential expression between SnCs and non-SnCs, and stronger associations with donor age in GTEx data.
Building on CoreScence, DeepScence employs a zero-inflated negative binomial (ZINB) autoencoder with a two-neuron bottleneck layer: one neuron captures senescence-related signal, the other captures unrelated variation. This design isolates senescence information without supervised labels, making it broadly applicable. The resulting continuous senescence score can optionally be binarized via a permutation-based procedure.
Across six in vitro scRNA-seq datasets, DeepScence achieved AUROCs exceeding 0.9 on all datasets and outperformed SenCID, AUCell, ssGSEA, and single-marker approaches. Critically, DeepScence also excelled on in vivo datasets — where SenCID, trained only on in vitro bulk data, showed markedly reduced performance. The method further demonstrated strong generalization to spatial transcriptomics data from platforms including 10× Visium and 10× Xenium, correctly localizing SnC-enriched regions consistent with histological annotations. DeepScence also generalized across mouse and human tissues and across multiple senescence induction contexts (replicative, oncogene-induced, radiation-induced, and drug-induced senescence).
These results establish DeepScence as a broadly applicable, platform-agnostic tool for studying where and how senescent cells accumulate in aging tissues and disease states. An important caveat is that ground-truth senescence labels in vivo remain imperfect, and the CoreScence gene set, while robust, may not capture all tissue-specific senescence programs. The tool is available as open-source software.
Key Findings
- Only 39 of 2,966 senescence-related genes were shared across five or more published gene sets, revealing massive inconsistency in the field.
- CoreScence, a 39-gene consensus set, showed stronger differential expression and age-association signals than genes from any single gene set.
- DeepScence achieved AUROCs >0.9 across all six in vitro scRNA-seq datasets, outperforming SenCID, AUCell, ssGSEA, and single-marker methods.
- Unlike SenCID, DeepScence generalized effectively to in vivo datasets and spatial transcriptomics platforms including 10× Visium and Xenium.
- DeepScence correctly mapped senescent cell-enriched spatial regions across tissues, species, and multiple senescence induction contexts.
Methodology
DeepScence is an unsupervised ZINB autoencoder trained on the 39-gene CoreScence set, evaluated against six in vitro and multiple in vivo scRNA-seq datasets plus spatial transcriptomics data from 10× Visium and Xenium platforms. Performance was measured by AUROC, accuracy, and F1 score, with paired t-tests comparing DeepScence to competing methods.
Study Limitations
Ground-truth senescence labels in in vivo datasets are imperfect, as no perfect single gold standard for SnC identity exists in complex tissues. CoreScence's 39-gene consensus set may miss tissue-specific or context-specific senescence programs not captured across multiple published gene sets. The model was validated primarily in human and mouse data, and generalizability to other species or highly sparse ST platforms requires further testing.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: